INTRODUCTION

Increasing use of prostate MRI in men with a clinically suspected or diagnosed prostate cancer (PCa) is placing increasing financial burden on healthcare systems worldwide (1). Biparametric MRI (bpMRI) (no iv contrast injection) has lower acquisition time and lower cost than routine multiparametric MRI. Moreover, further reduction of scan time can be achieved by careful optimization of bpMRI acquisition protocol. In the current study, we aimed to develop, validate and test an open-source machine learning tools for bpMRI with acquisition time of only 5 minutes (Figure 1).

METHODS

One hundred fifty-seven men with diagnosed PCa underwent 3T prostate MR (Ingenuity PET/ MR, Philips, Best, the Netherlands) using 32 channel manufacturer's cardiac coils before prostatectomy. Transversal T2-weighted images (T2w) were acquired using a single-shot turbo spin-echo sequence with TR/TE 4668/130 ms, FOV 250x320 mm², acquisition/reconstruction matrix size 250x320/512x672, slice thickness 2.5 mm and scan time of 1 minute 10 seconds. Diffusion weighted imaging was performed with single-shot SE epi sequence with TR/TE 3141 ms/51 ms, FOV 250×250 mm², acquisition/reconstruction matrix 100x99 / 224x224, slice thickness 5 mm, diffusion gradient timing (Δ) 24.5 ms, diffusion gradient duration (δ) 12.6 ms, 4 of scanned 12 b-values were used (number of signal averages) 0 (2), 900 (2), 1100 (2), 2000 (4) s/mm² (2), and scan time of 3 minutes and 17 seconds. The combined scan time of T2w and DWI (4 b-values) was 4 minutes and 27 seconds.

DWI data sets were fitted using monoexponential (ADCm) function utilizing the Broyden-Fletcher-Goldfarb-Shanno algorithm in Leastsqbound‐scipy library (3). T2w signal intensities (“intensity drift”) were corrected using histogram alignment method as described by Nyul et al. (4). Whole prostate gland was manually delineated and the volumes were considered for histogram alignment (Figure 1). Each individual tumor focus in whole mount prostatectomy sections was graded separately based on the International Society of Urological Pathology (ISUP) guidelines (5) and Gleason Grade Groups (GGG) were assigned. Prostate cancer extent on each MRI acquisition (T2w, DWI) was manually delineated using whole mount prostatectomy sections as “ground truth”.

Feature extraction
First-order statistics (mean, standard deviation, median, range, 25th and 75th percentile) from T2w and ADCm signal intensities were calculated. Additionally, features derived from edge detector maps (6) (Figure 1), shape features, high-pass and low-pass filtered data, and 3D Laws features were obtained. Low and high-pass filters were implemented before extracting statistics from signal to evaluate the performance of other implemented feature extraction techniques in relation to simple noise filtering to the data. Finally, a three-dimensional version of Laws features (7) and shape feature of surface curvature (8, 9) were extracted from 3D mesh after applying marching cubes algorithm to the binary lesion. All 3D feature extractions were preceded by resampling image data into isotropic voxel size. When applicable, all features were calculated from whole prostate, PCa lesions, and whole gland region excluding the lesion. Multiple parameters were applied to feature extraction algorithms where considered suitable.

Feature selection and machine learning
Random stratified (based on proportion of low grade GGG 1,2 vs >2) split with 70-30 ratio was performed into training (n=107) and unseen testing (n=50) data (Figure 2, 3). In total, 5836 features which have earlier demonstrated good short-term repeatability (10), were calculated from T2w and ADCm. The features were grouped by their functioning principle, and 2 best features were included from each of 9 feature groups using Minimum Redundancy Maximum Relevance (MRMR), followed by Nearest Neighbour based relevance calculation using Boruta R package (11), removing features expressing significantly low relevance. Finally, clusters found significant with Pvclust package (12) were reduced by removing redundant features until no features were found to resemble others, resulting 9, 8 and 4 features for feature sets from T2w only, ADCm only, and T2w combined with ADCm, respectively. The training data of 107 patients was used in 3-fold cross validation to estimate performance of three logistic regressor models for T2w, ADCm and T2w combined with ADCm. The performance measure was Area Under Receiver Operator Characteristic Curve (AUC) for prediction of GGG 1 vs >1 based on whole mount prostatectomy. Feature selection, regress or fitting and statistical analyses were implemented using R (version 3.5.3, R Foundation for Statistical Computing, Vienna, Austria).

RESULTS

Best individual features selected with training set were ADCm Frangi filter intensity, and T2w Harris-Stephens number of corner locations ratio (Figure 4). Their performance in unseen data in terms of AUC (95%CI) was 0.82 (0.63-0.96) and 0.85 (0.75-0.93) in lesion-wise analysis. Features selected for logistic regressor are listed in Figure 4. Logistic regressor classification performance AUC (95%CI) was 0.86 (0.77-0.94), 0.89 (0.81-0.96), 0.91 (0.82-0.98) for T2w alone, ADCm alone and bpMRI T2W with ADCm, respectively.

DISCUSSION/CONCLUSION

Machine learning using radiomic/texture features extracted from optimized 5-minute bpMRI (2) have a good classification performance for differentiating PCa with GGG≤1 (“clinically insignificant PCa”) versus PCa with GGG>1 (“clinically significant PCa”).

Acknowledgements

This study was financially supported by grants from Orion Pharma Research Foundation (HM), Sigrid Jusélius Foundation (HM, HA, IJ), Finnish Cultural Foundation (HM, IJ), clinical Researcher Funding from the Academy of Finland (PT), TYKS-SAPA Research funds (HM, PT,OE, MP, JS, JPB, HA, IJ), Finnish Cancer Society (IJ), Instrumentarium Research Foundation (IJ, HA). Funding from Instrumentarium Science Foundation, Sigrid Jusélius Foundation, Turku University Hospital, TYKS-SAPA research funds were used to cover the cost of MRI examinations.